Computer vision has become indispensable in various applications, including autonomous driving, medical imaging, security and surveillance, robotics, and pattern recognition. In recent years, the quality of training data has emerged as a critical factor for ensuring effectiveness in real-world scenarios. However, the increasing stringency of privacy regulations in various regions necessitates careful handling of collected images for computer vision. Personal information within images is typically anonymized by applying anonymization patterns to remove it. Empirical findings underscore the significant influence of data quality on the training of deep learning models. Striking the right balance between privacy and recognition performance becomes paramount. Therefore, it is essential to understand how the anonymization of image datasets affects deep learning model performance. In this paper, we thoroughly analyze the effects of different anonymization techniques on the performance of deep learning-based models in computer vision tasks, with a particular emphasis on presenting a model-centric perspective, such as the type of deep learning model and the number of parameters. We aim to provide valuable insights and guidelines for selecting the optimal level of anonymization that strikes a balance between recognition accuracy and privacy protection.

Human Activity Recognition (HAR) is a challenging task in the field of human-related signal processing. Owing to the development of wearable sensing technology, an emerging research approach in HAR is to identify user-performed tasks by using data collected from wearable sensors. In this paper, we propose a novel system for monitoring and recognizing daily living activities using an off-the-shelf smart band and two smart shoes. The system aims at providing a useful tool for solving problems regarding body part placement, fusion of multimodal sensors and feature selection for a specific set of activities. The system collects inertial and plantar pressure data at wrist and foot to analyze and then, extract, select important features for recognition. We construct and compare two predictive models of classifying activities from the reduced feature set. A comparison of the classification for each wearable device and a fusion scheme is provided to identify the best body part for activity recognition: either the wrist or the feet. This comparison also demonstrated the effective HAR performance of the proposed system.

Advanced research in robotics has allowed robots to navigate diverse environments autonomously. However, conducting complex tasks while handling unpredictable circumstances is still challenging for robots. The robots should plan the task by understanding the working environments beyond metric information and need countermeasures against various situations. In this paper, we propose a semantic navigation framework based on a Triplet Ontological Semantic Model (TOSM) to manage various conditions affecting the execution of tasks. The framework allows robots with different kinematics to perform tasks in indoor and outdoor environments. We define the TOSM-based semantic knowledge and generate a semantic map for the domains. The robots execute tasks according to their characteristics by converting inferred knowledge to Planning Domain Definition Language (PDDL). Additionally, to make the framework sustainable, we determine a policy of maintaining the map and re-planning when in unexpected situations. The various experiments on four different kinds of robots and four scenarios validate the scalability and reliability of the proposed framework.

In this paper, we propose a novel method for ambulatory activity recognition and pedestrian identification based on temporally adaptive weighting accumulation-based features extracted from categorical plantar pressure. The method relies on three pressure-related features, which are calculated by accumulating the pressure of the standing foot in each step over three different temporal weighting forms. In addition, we consider a feature reflecting the pressure variation. These four features characterize the standing posture in a step by differently weighting step pressure data over time. We use these features to analyze the standing foot during walking and then recognize ambulatory activities and identify pedestrians based on multilayer multiclass support vector machine classifiers. Experimental results show that the proposed method achieves 97% accuracy for the two tasks when analyzing eight consecutive steps. For faster processing, the method reaches 89.9% and 91.3% accuracy for ambulatory activity recognition and pedestrian identification considering two consecutive steps, respectively, whereas the accuracy drops to 83.3% and 82.3% when considering one step for the respective tasks. Comparative results demonstrated the high performance of the proposed method regarding accuracy and temporal sensitivity.

Understanding human hand usage is one of the richest information source to recognize human manipulation actions. Since humans use various tools during actions, grasp recognition gives important cues to figure out humans' intention and tasks. Earlier studies analyzed grasps with positions of hand joints by attaching sensors, but since these types of sensors prevent humans from naturally conducting actions, visual approaches have been focused in recent years. Convolutional neural networks require a vast annotated dataset, but, to our knowledge, no human grasping dataset includes ground truth of hand regions. In this paper, we propose a grasp recognition method only with image-level labels by the weakly supervised learning framework. In addition, we split the grasp recognition process into two stages that are hand localization and grasp classification so as to speed up. Experimental results demonstrate that the proposed method outperforms existing methods and can perform in real-time.

Numerical weather prediction (NWP) models are fundamental in meteorology for simulating and forecasting the behavior of various atmospheric variables. The accuracy of precipitation forecasts and the acquisition of sufficient lead time are crucial for preventing hazardous weather events. However, the performance of NWP models is limited by the nonlinear and unpredictable patterns of extreme weather phenomena driven by temporal dynamics. In this regard, we propose a \textbf{S}elf-\textbf{S}upervised \textbf{L}earning with \textbf{P}robabilistic \textbf{D}ensity \textbf{L}abeling (SSLPDL) for estimating rainfall probability by post-processing NWP forecasts. Our post-processing method uses self-supervised learning (SSL) with masked modeling for reconstructing atmospheric physics variables, enabling the model to learn the dependency between variables. The pre-trained encoder is then utilized in transfer learning to a precipitation segmentation task. Furthermore, we introduce a straightforward labeling approach based on probability density to address the class imbalance in extreme weather phenomena like heavy rain events. Experimental results show that SSLPDL surpasses other precipitation forecasting models in regional precipitation post-processing and demonstrates competitive performance in extending forecast lead times. Our code is available at https://github.com/joonha425/SSLPDL